AI Fitness OS: Verifiable, Multi-Metric Fitness Guidance by Constraining a Large Language Model with Deterministically Computed Targets

Abstract

Fitness applications usually record one behaviour in isolation, and large-language-model (LLM) coaches produce fluent advice whose numbers are rarely checked. This paper presents AI Fitness OS, a web application that keeps training, nutrition, hydration, sleep, weight, steps and goals in one per-user record, computes every numerical target deterministically from published equations and stated heuristics, and passes those targets to Google Gemini as binding constraints; each generated meal plan is re-checked in code against a ±5 % tolerance. The premise was examined through secondary research on a public dataset of 2,111 records, and the artefact through tests, simulation and live measurement. All eight lifestyle metrics that the system tracks or addresses were significantly associated with obesity level (Bonferroni-adjusted ? = 0.00625), although only three remained so on 515 records resembling the original survey. A Random Forest that combined the eight metrics reached a macro-F1 of 0.720 against 0.263 for the best single metric (Wilcoxon W = 55, p = 0.00098). Across 1,000 simulated user profiles, every constrained meal plan fell within ±5 % of its calorie target, against 75 of 1,000 unconstrained plans (exact McNemar p < 10?²??). The calculator agreed with the dataset\'s weight-category labels for 98.82 % of records, and live responses took 2.5–6.1 s at no cost. The findings support computing arithmetic, generating language and checking the output against the computed numbers.

Introduction

AI Fitness OS is an integrated, AI-assisted fitness management platform designed to overcome the problem of fragmented health tracking. It combines physical activity, diet, water intake, sleep, weight, steps, and fitness goals in a single system to provide personalized health recommendations.

The system uses deterministic calculations based on established equations to calculate Body Mass Index (BMI), Basal Metabolic Rate (BMR), daily energy expenditure, calorie requirements, protein, fat, carbohydrate, and water targets. Google Gemini AI generates personalized workout plans, Indian-style meal plans, healthier food alternatives, daily tips, and coaching responses based on these calculated targets. Generated meal plans are also checked programmatically, with portion sizes adjusted when calorie totals differ from the target by more than 5%.

The platform is developed using Flutter Web, Riverpod, Firebase Authentication, and Cloud Firestore, with the aim of providing a secure, responsive, and zero-cost fitness management solution.

The study evaluates the system using public obesity-related data, statistical analysis, automated testing, simulated user profiles, calculator validation, and live AI response measurements. The dataset contains 2,111 records, and the BMI calculator achieved 98.82% agreement with the dataset's obesity categories after classification mapping. Statistical analysis found significant associations between all eight lifestyle metrics and obesity levels in the full dataset, although these associations were weaker in the subset of original-like records, highlighting the limitations of synthetic data.

The main objective is to develop a reliable and reproducible fitness platform that integrates multiple health indicators, generates personalized AI recommendations, verifies numerical accuracy, and encourages healthier lifestyle habits. Unlike conventional fitness applications, AI Fitness OS combines health tracking, scientifically based calculations, AI-generated guidance, and programmatic verification in one unified system.

Conclusion

This study presented AI Fitness OS, a zero-cost web application that unifies a user\'s training, nutrition, hydration, sleep, weight and step data, computes all targets deterministically from published equations and stated heuristics, and constrains Google Gemini with those targets while verifying each meal plan in code. Secondary research on 2,111 records showed that the metrics the system tracks or addresses are associated with obesity level and that, combined, they predict it far better than any single metric; a 1,000-profile simulation showed that the constraint-and-check mechanism places every meal plan within ±5 % of its calorie target, against 7.5 % without it. The conclusions are bounded by honest limitations: about 77 % of the dataset is synthetic, the respondents come from three Latin-American countries and answered by self-report, only two live LLM meal plans were measured, and no study with users was conducted. Within those bounds, the work shows that a large language model can be made verifiable in a health setting by a simple rule — arithmetic is computed, language is generated, and the generated output is checked against the computed numbers.

References

[1] Z. He, M. A. Hassan, P. Saiz-González, S. Ryu, R. Wang, and Z. Gao, “Smartphone app-based interventions on physical activity behaviors and psychological correlates in healthy young adults: A systematic review,” PLOS ONE, vol. 19, no. 4, p. e0301088, 2024. [2] H. Lin, M. Yang, Z. Zhou, Y. Zhang, and N. Deng, “Exploring the dynamics of dietary self-monitoring adherence among participants in a digital behavioral weight loss program: Model development study,” J. Med. Internet Res., vol. 27, p. e65431, 2025. [3] D. Shin, G. Hsieh, and Y.-H. Kim, “PlanFitting: Personalized exercise planning with large language model-driven conversational agent,” arXiv:2309.12555, 2023. [4] Z. Yang et al., “ChatDiet: Empowering personalized nutrition-oriented food recommender chatbots through an LLM-augmented framework,” arXiv:2403.00781, 2024. [5] World Health Organization, “Obesity: Preventing and managing the global epidemic,” WHO Technical Report Series 894, Geneva, 2000. [6] M. D. Mifflin, S. T. St Jeor, L. A. Hill, B. J. Scott, S. A. Daugherty, and Y. O. Koh, “A new predictive equation for resting energy expenditure in healthy individuals,” Am. J. Clin. Nutr., vol. 51, no. 2, pp. 241–247, 1990. [7] M. Jörke et al., “GPTCoach: Towards LLM-based physical activity coaching,” arXiv:2405.06061, 2024. [8] J. Cosentino, A. Belyaeva, X. Liu, and N. A. Furlotte, “Towards a personal health large language model,” arXiv:2406.06474, 2024. [9] A. A. Heydari, K. Gu, V. Srinivas, and H. Yu, “The anatomy of a personal health agent,” arXiv:2508.20148, 2025. [10] S. Khamesian, A. Arefeen, S. M. Carpenter, and H. Ghasemzadeh, “NutriGen: Personalized meal plan generator leveraging large language models to enhance dietary and nutritional adherence,” arXiv:2502.20601, 2025. [11] F. Gao et al., “HealthGenie: Empowering users with healthy dietary guidance through knowledge graph and large language models,” arXiv:2504.14594, 2025. [12] I. Azimi, M. Qi, L. Wang, A. M. Rahmani, and Y. Li, “Accuracy and consistency of LLMs in the registered dietitian exam: The impact of prompt engineering and knowledge retrieval,” arXiv:2408.02964, 2024. [13] X. Lai, Y. Lai, J. Chen, S. Huang, Q. Gao, and C. Huang, “Evaluation strategies for large language model-based models in exercise and health coaching: Scoping review,” J. Med. Internet Res., vol. 27, p. e79217, 2025. [14] Y. Sun, H. Dong, and W. Jiang, “Is more always better? An S-shaped impact of gamification feature richness on exercise adherence intention,” Front. Psychol., vol. 16, art. 1671543, 2025. [15] J. W. Navalta, J. D. Thomas, and W. J. Stone, “Demographic profiles and methodologies used in the generation and validation of resting metabolic rate prediction equations: Protocol for a systematic review,” JMIR Res. Protoc., vol. 15, p. e82482, 2026. [16] J. D. Stookey et al., “Hydration and health at ages 40–70 years in Salzburg Austria is associated with a median total water intake over 40 mL/kg including at least 1 L/d plain drinking water,” Front. Public Health, vol. 13, art. 1668981, 2025. [17] A. R. Hevner, S. T. March, J. Park, and S. Ram, “Design science in information systems research,” MIS Quarterly, vol. 28, no. 1, pp. 75–105, 2004. [18] F. M. Palechor and A. de la Hoz Manotas, “Dataset for estimation of obesity levels based on eating habits and physical condition in individuals from Colombia, Peru and Mexico,” Data in Brief, vol. 25, art. 104344, 2019. [19] N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, “SMOTE: Synthetic minority over-sampling technique,” J. Artif. Intell. Res., vol. 16, pp. 321–357, 2002. [20] F. Pedregosa et al., “Scikit-learn: Machine learning in Python,” J. Mach. Learn. Res., vol. 12, pp. 2825–2830, 2011. [21] W. G. Cochran, Sampling Techniques, 3rd ed. New York: Wiley, 1977. [22] W. H. Kruskal and W. A. Wallis, “Use of ranks in one-criterion variance analysis,” J. Amer. Statist. Assoc., vol. 47, no. 260, pp. 583–621, 1952. [23] L. Breiman, “Random forests,” Machine Learning, vol. 45, no. 1, pp. 5–32, 2001. [24] F. Wilcoxon, “Individual comparisons by ranking methods,” Biometrics Bull., vol. 1, no. 6, pp. 80–83, 1945.

Copyright

Copyright © 2026 Selvin Lopes, Perseus Bhavnagri. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.